AI Spots Melanoma On Light Skin. Dark Skin Gets Missed. The Training Data Is The Problem, Obviously.

AI skin cancer detection tools are being trained on medical image libraries dominated by lighter skin tones, creating a systematic blind spot for darker skin. A harmless dark spot on light skin may trigger false panic, while a life-threatening cancer on dark skin could be overlooked entirely. Researchers are now exploring whether text prompts can generate realistic images of conditions like melanoma on darker skin to rebalance the training data.

This is a textbook case of representation bias, a term I suspect most of you have not encountered. The principle is straightforward: a model trained on a non-representative sample will produce non-representative outputs. If your training data is 90 percent light skin, your model is not an AI dermatologist. It is an AI dermatologist for light skin. The mechanism is called sampling bias, and it is the single most underappreciated failure mode in applied machine learning.

Researchers building these AI models are pulling hundreds of thousands of images from public online libraries shared by universities and hospitals. The same bias exists in dermatology textbooks, which have historically been dominated by lighter skin tones.